Re: [PATCH] route:ip_rt_frag_needed always return unzero

[Eric Dumazet <eric.dumazet@gmail.com>]

Le mercredi 19 octobre 2011 à 09:34 +0800, Gao feng a écrit :
> 2011.10.18 17:23, Eric Dumazet wrote:
> > Le mardi 18 octobre 2011 à 15:04 +0800, Gao feng a écrit :
> >> int function ip_rt_frag_need,if peer is null,
> >> there is no need to do ipprot->err_handler.
> >> I am right?
> >>
> >> Signed-off-by: Gao feng <gaofeng@cn.fujitsu.com>
> >> ---
> >>  net/ipv4/route.c |    2 +-
> >>  1 files changed, 1 insertions(+), 1 deletions(-)
> >>
> >> diff --git a/net/ipv4/route.c b/net/ipv4/route.c
> >> index 075212e..6cde0fa 100644
> >> --- a/net/ipv4/route.c
> >> +++ b/net/ipv4/route.c
> >> @@ -1574,7 +1574,7 @@ unsigned short ip_rt_frag_needed(struct net *net, const struct iphdr *iph,
> >>  
> >>  		atomic_inc(&__rt_peer_genid);
> >>  	}
> >> -	return est_mtu ? : new_mtu;
> >> +	return est_mtu;
> >>  }
> >>  
> >>  static void check_peer_pmtu(struct dst_entry *dst, struct inet_peer *peer)
> > 
> > No idea why you want this, your changelog is a bit cryptic :)
> > 
> > Wont this bypass the raw_icmp_error(skb, protocol, info);
> > call in icmp_unreach() as well ?
> > 
> > 
> 
> thanks Eric!
> 
> I mean that the pmtu is update by inet_peer->pmtu_learned as I know.
> so in function ip_rt_frag_needed,
> if inet_peer is null or someting else make the setting of inet_peer->pmtu_learned failed.
> there is no need to call function tcp_v4_err.
> 
> the call stack is
> icmp_unreach
>   |
>   |--->ip_rt_frag_needed(fill inet_peer)
>   |
>   |--->raw_icmp_error()
>   |
>   |--->ipprot->err_handler(tcp_v4_err or something else)
> 	|
> 	|--->tcp_v4_err(frag need icmp is triggered by tcp packet)
> 		|
> 		|--->do_pmtu_discovery
> 		(in this function both __sk_dst_check or dst->ops->update_pmtu
> 		need struct inet_peer to update pmtu)
> 
> so,I think when set inet_peer->pmtu_learned failed,
> in func icmp_unreach we should goto out immediately.
> 
> And it's confuse me that why func ping_err and udp_err not update the pmtu?
> What I miss?

You dont answer my question : After your patch, we now dont call
raw_icmp_error() anymore. Why is is valid ?

Not finding/create inet_peer is very unlikely : This occurs only under
high stress and out of memory condition. Is it really happening on your
machines ?